Baseline GBM model and Improved XGB model

In your final repo, there should be an R markdown file that organizes all computational steps for evaluating your proposed Facial Expression Recognition framework.

This file is currently a template for running evaluation experiments. You should update it according to your codes but following precisely the same structure.

Step 0 set work directories

set.seed(5)

Provide directories for training images. Training images and Training fiducial points will be in different subfolders.

train_dir <- ("../data/train_set/") # This will be modified for different data sets.
train_image_dir <- paste(train_dir, "images/", sep="")
train_pt_dir <- paste(train_dir,  "points/", sep="")
train_label_path <- paste(train_dir, "label.csv", sep="") 

Step 1: set up controls for evaluation experiments.

In this chunk, we have a set of controls for the evaluation experiments.

  • (T/F) fit gbm model on the training set
  • (T/F) run evaluation on an independent test set
  • (T/F) process features for training set
  • (T/F) process features for test set
run.cv <- TRUE # run cross-validation on the training set
sample.reweight <- TRUE # run sample reweighting in model training
smote <- TRUE # run SMOTE on in model training
K <- 5  # number of CV folds
run.gbm = TRUE # run the gbm model on the training set
run.test.gbm = TRUE # run evaluation on an independent test set
run.feature.train <- TRUE # process features for training set
run.feature.test <- TRUE # process features for test set
run.test <- TRUE # run evaluation on an independent test set

Using cross-validation or independent test set evaluation, we compare the performance of models with different specifications. In this Starter Code, we tune parameter k (number of neighbours) for KNN.

Step 2: import data and train-test split

#train-test split
info <- read.csv("../data/train_set/label.csv")
n <- nrow(info)
n_train <- round(n*(4/5), 0)
train_idx <- sample(info$Index, n_train, replace = F)
test_idx <- setdiff(info$Index,train_idx)

If you choose to extract features from images, such as using Gabor filter, R memory will exhaust all images are read together. The solution is to repeat reading a smaller batch(e.g 100) and process them.

n_files <- length(list.files(train_image_dir))
image_list <- list()
for(i in 1:100){
   image_list[[i]] <- readImage(paste0(train_image_dir, sprintf("%04d", i), ".jpg"))
}

Fiducial points are stored in matlab format. In this step, we read them and store them in a list.

#function to read fiducial points
#input: index
#output: matrix of fiducial points corresponding to the index
readMat.matrix <- function(index){
     return(round(readMat(paste0(train_pt_dir, sprintf("%04d", index), ".mat"))[[1]],0))
}
#load fiducial points
fiducial_pt_list <- lapply(1:n_files, readMat.matrix)
save(fiducial_pt_list, file="../output/fiducial_pt_list.RData")

Step 3: construct features and responses

  • The follow plots show how pairwise distance between fiducial points can work as feature for facial emotion recognition.

    • In the first column, 78 fiducials points of each emotion are marked in order.
    • In the second column distributions of vertical distance between right pupil(1) and right brow peak(21) are shown in histograms. For example, the distance of an angry face tends to be shorter than that of a surprised face.
    • The third column is the distributions of vertical distances between right mouth corner(50) and the midpoint of the upper lip(52). For example, the distance of an happy face tends to be shorter than that of a sad face.
Figure1

Figure1

feature.R should be the wrapper for all your feature engineering functions and options. The function feature( ) should have options that correspond to different scenarios for your project and produces an R object that contains features and responses that are required by all the models you are going to evaluate later.

  • feature.R
  • Input: list of images or fiducial point
  • Output: an RData file that contains extracted features and corresponding responses
source("../lib/feature.R")
tm_feature_train <- NA
if(run.feature.train){
  tm_feature_train <- system.time(dat_train <- feature(fiducial_pt_list, train_idx))
}
#save(dat_train, file="../output/feature_train.RData")

tm_feature_test <- NA
if(run.feature.test){
  tm_feature_test <- system.time(dat_test <- feature(fiducial_pt_list, test_idx))
}
#save(dat_test, file="../output/feature_test.RData")

Baseline Model: GBM

load lib files

source("../lib/cross_validation_gbm.R")
source("../lib/train_gbm.R")
source("../lib/test_gbm.R")
source("../lib/feature.R")

Model selection with cross-validation

  • Do model selection by choosing among different values of training model parameters.
tm_train_gbm_baseline <- NA
if(run.gbm){
  # Train the Baseline GBM model
  tm_train_gbm_baseline <- system.time(gbm.baseline <- train_gbm(train_data = dat_train,  s=0.001, K=2, n=50))
  # Save the output
  save(gbm.baseline, file="../output/gbm.baseline")
}

tm_train_gbm_baseline <- system.time(gbm.baseline <- train_gbm(train_data = dat_train,  s=0.001, K=2, n=50))

Step 5: Run test on test images

load('../output/gbm.baseline')
run.test = TRUE
tm_test = NA
tm_test_gbm_baseline <- NA
if(run.test.gbm){
  tm_test_gbm_baseline <- system.time(pred_gbm_baseline <- test_gbm(gbm.fit.model = gbm.baseline, input.test = dat_test[,-6007], n = 50))
}

tm_test_gbm_baseline <- system.time(pred_gbm_baseline <- test_gbm(gbm.fit.model = gbm.baseline, input.test = dat_test[,-6007], n = 50))
  • evaluation
accuracy_baseline_gbm <- mean(dat_test$label == pred_gbm_baseline)
pred_gbm_baseline_num <- as.numeric(pred_gbm_baseline)
tpr.fpr <- WeightedROC(pred_gbm_baseline_num, dat_test$label)
auc <- WeightedAUC(tpr.fpr)
cat("The accuracy of model: GBM baseline is", mean(dat_test$label == pred_gbm_baseline)*100, "%.\n")
## The accuracy of model: GBM baseline is 79.5 %.
cat("The AUC of model: GBM baseline is", auc, ".\n")
## The AUC of model: GBM baseline is 0.5 .
# Confusion Matrix 
library(caret)
# confusionMatrix(dat_test$label, as.factor(pred_gbm_baseline))

Summarize Running Time

Prediction performance matters, so does the running times for constructing features and for training the model, especially when the computation resource is limited.

# Model_performace <- function(time_feature_train, time_feature_test, time_train, time_test){
#   cat("Time for constructing training features=", time_feature_train[1], "s \n")
#   cat("Time for constructing testing features=", time_feature_test[1], "s \n")
#   cat("Time for training model=", time_train[1], "s \n") 
#   cat("Time for testing model=", time_test[1], "s \n")
# } 
tm_train_gbm <- 72.382
tm_test_gbm <- 18.267
print(paste("Time for constructing training features=", tm_feature_train,  "s"))
## [1] "Time for constructing training features= 0.876000000000001 s"
## [2] "Time for constructing training features= 0.339 s"            
## [3] "Time for constructing training features= 1.224 s"            
## [4] "Time for constructing training features= 0 s"                
## [5] "Time for constructing training features= 0 s"
print(paste("Time for constructing testing features=", tm_feature_test, "s"))
## [1] "Time for constructing testing features= 0.187000000000001 s" 
## [2] "Time for constructing testing features= 0.0629999999999997 s"
## [3] "Time for constructing testing features= 0.251000000000001 s" 
## [4] "Time for constructing testing features= 0 s"                 
## [5] "Time for constructing testing features= 0 s"
print(paste("Time for training model=", tm_train_gbm, "s"))
## [1] "Time for training model= 72.382 s"
print(paste("Time for testing model=", tm_test_gbm, "s"))
## [1] "Time for testing model= 18.267 s"

Improved Model: XGboosting

Step 4: Train a classification model with training features and responses

source("../lib/train_xgb.R") 
source("../lib/test_xgb.R")
# cross validation with reweighting data, tuning nrounds and max_depth
source("../lib/cross_validation_xgb.R")

Model selection with cross-validation

Cross Validation with Reweighting sample
nrounds_list <- seq(20, 100, 20)
max_depth_list <- c(10, 20)
run.cv <- FALSE
if(run.cv){
  res_cv_rw <- cv.function.xgb(dat_train, 5, reweight = TRUE, smote = FALSE, 
                               nrounds_list, max_depth_list)
  save(res_cv_rw, file="../output/res_cv_rw.RData")
}else{
  load("../output/res_cv_rw.RData")
}

res_cv_rw
## $mean_error
##           [,1]      [,2]
## [1,] 0.3433302 0.3636274
## [2,] 0.3565976 0.3589426
## [3,] 0.3707250 0.3487006
## [4,] 0.3597258 0.3439876
## [5,] 0.3427746 0.3422774
## 
## $mean_AUC
##           [,1]      [,2]
## [1,] 0.7837712 0.7768042
## [2,] 0.7899824 0.7978492
## [3,] 0.8002050 0.8092118
## [4,] 0.8041076 0.8001706
## [5,] 0.8012920 0.8074034
Cross Validation with SMOTE
# cross validation with SMOTE, tuning nrounds and max_depth
source("../lib/cross_validation.R")

nrounds_list <- seq(20, 100, 20)
max_depth_list <- c(10, 20)
run.cv <- FALSE
if(run.cv){
  res_cv_sm <- cv.function.xgb(dat_train, 5, reweight = FALSE, smote = TRUE, 
                               nrounds_list, max_depth_list)
  save(res_cv_sm, file="../output/res_cv_sm.RData")
}else{
  load("../output/res_cv_sm.RData")
}

res_cv_sm
## $mean_error
##           [,1]      [,2]
## [1,] 0.0921242 0.0959984
## [2,] 0.0963390 0.0798202
## [3,] 0.0847360 0.0843836
## [4,] 0.0826272 0.0840486
## [5,] 0.0794810 0.0872006
## 
## $mean_AUC
##           [,1]      [,2]
## [1,] 0.9732210 0.9693970
## [2,] 0.9742660 0.9765240
## [3,] 0.9776792 0.9777940
## [4,] 0.9801844 0.9762790
## [5,] 0.9785240 0.9765744
  • Train the model with the entire training set using the selected model (model parameter) via cross-validation.

Reweighting sample

# source("../lib/train.R") 
train_label <- as.numeric(levels(dat_train$label))[dat_train$label]
weight_train <- rep(NA, length(train_label))
for (v in unique(train_label)){
  weight_train[train_label == v] = 0.5 * length(train_label) / length(train_label[train_label == v])
}


train_xgb <- xgb.DMatrix(as.matrix(dat_train[, -6007]), 
                           label = train_label,
                           weight = weight_train)
tm_train_rw <- system.time(fit_train_rw <- train.xgb(train_xgb, nrounds = 100, max_depth = 20))

save(fit_train_rw, file="../output/fit_train_rw.RData")
# oversampling data using SMOTE method
library(DMwR)
set.seed(2020)
tm_smote <- system.time(train_smote <- SMOTE(label ~ ., dat_train, perc.over = 200, k = 5, 
                                             perc.under = 150))
tm_smote
##    user  system elapsed 
##  50.939  21.076  72.334

SMOTE

train_label <- as.numeric(levels(train_smote$label))[train_smote$label]
train_xgb <- xgb.DMatrix(as.matrix(train_smote[, -6007]), label = train_label)

tm_train_sm <- system.time(fit_train_sm <- train.xgb(train_xgb, nrounds = 80, max_depth = 10))

save(fit_train_sm, file="../output/fit_train_sm.RData")

Step 5: Run test on test images

Reweighting sample

test_label <- as.numeric(levels(dat_test$label))[dat_test$label]
weight_test <- rep(NA, length(test_label))
for (v in unique(test_label)){
  weight_test[test_label == v] = 0.5 * length(test_label) / length(test_label[test_label == v])
}

tm_test_rw = NA
feature_test <- as.matrix(dat_test[, -6007])
if(run.test){
  load(file="../output/fit_train_rw.RData")
  test_label <- as.numeric(levels(dat_test$label))[dat_test$label]
  test_xgb <- xgb.DMatrix(as.matrix(dat_test[, -6007]), 
                            label = test_label,
                            weight = weight_test)
  tm_test_rw <- system.time({prob_pred <- test.xgb(fit_train_rw, test_xgb)[1];
                          label_pred <- test.xgb(fit_train_rw, test_xgb)[2]})
}
  • evaluation
## reweight the test data to represent a balanced label distribution
label_test <- as.integer(dat_test$label)-1
weight_test <- rep(NA, length(label_test))
for (v in unique(label_test)){
  weight_test[label_test == v] = 0.5 * length(label_test) / length(label_test[label_test == v])
}

accu <- sum(weight_test * as.numeric(unlist(label_pred)) == label_test)/sum(weight_test)
tpr.fpr <- WeightedROC(as.numeric(unlist(prob_pred)), label_test, weight_test)
auc <- WeightedAUC(tpr.fpr)


cat("The accuracy of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is", accu*100, "%.\n")
cat("The AUC of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is", auc, ".\n")
## The accuracy of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is 75.83333 %.
## The AUC of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is 0.8494486 .
  • confusion matrix
label_pred <- as.numeric(unlist(label_pred))
cf_mat <- table(label_pred, label_test)
cf_mat

TN <- cf_mat[1, 1]
FP <- cf_mat[2, 1]
FN <- cf_mat[1, 2]
TP <- cf_mat[2, 2]

Precision <- TP/(TP + FP)
Sensitivity <- TP/(TP + FN)
Specificity <- TN/(TN + FP)
F_score <- 2 * Precision * Sensitivity/(Precision + Sensitivity)

cat("The Precision of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is", 
    Precision * 100, "%.\n")
cat("The Sensitivity of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is", 
    Sensitivity * 100, "%.\n")
cat("The Specificity of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is", 
    Specificity * 100, "%.\n")
cat("The F score of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is", 
    F_score * 100, "%.\n")
##           label_test
## label_pred   0   1
##          0 455  67
##          1  22  56
## The Precision of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is 71.79487 %.
## The Sensitivity of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is 45.52846 %.
## The Specificity of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is 95.38784 %.
## The F score of model XGB with reweighting sample, nrounds = 100, max_depth = 20 is 55.72139 %.

SMOTE

# SMOTE
source("../lib/test.R")
tm_test = NA
feature_test <- as.matrix(dat_test[, -6007])
if(run.test){
  load(file="../output/fit_train_sm.RData")
  test_label <- as.numeric(levels(dat_test$label))[dat_test$label]
  test_xgb <- xgb.DMatrix(as.matrix(dat_test[, -6007]), 
                            label = test_label)
  tm_test_sm <- system.time({prob_pred <- test.xgb(fit_train_sm, test_xgb)[1];
                          label_pred <- test.xgb(fit_train_sm, test_xgb)[2]})
}
  • evaluation
# SMOTE
label_test <- as.integer(dat_test$label)-1


accu <- sum(as.numeric(unlist(label_pred)) == label_test)/length(label_test)
tpr.fpr <- WeightedROC(as.numeric(unlist(prob_pred)), label_test)
auc <- WeightedAUC(tpr.fpr)


cat("The accuracy of model XGB with SMOTE, nrounds = 80, max_depth = 10 is", accu*100, "%.\n")
cat("The AUC of model XGB with SMOTE, nrounds = 80, max_depth = 10 is", auc, ".\n")
## The accuracy of model XGB with SMOTE, nrounds = 80, max_depth = 10 is 81 %.
## The AUC of model XGB with SMOTE, nrounds = 80, max_depth = 10 is 0.8325578 .
  • confusion matrix
label_pred <- as.numeric(unlist(label_pred))
cf_mat <- table(label_pred, label_test)

TN <- cf_mat[1,1]
FP <- cf_mat[2,1]
FN <- cf_mat[1,2]
TP <- cf_mat[2,2]

Precision <- TP/(TP+FP)
Sensitivity <- TP/(TP+FN)
Specificity <- TN/(TN+FP)
F_score <- 2*Precision*Sensitivity/(Precision+Sensitivity)

cat("The Precision of model XGB with SMOTE, nrounds = 100, max_depth = 20 is", Precision*100, "%.\n")
cat("The Sensitivity of model XGB with SMOTE, nrounds = 100, max_depth = 20 is", Sensitivity*100, "%.\n")
cat("The Specificity of model XGB with SMOTE, nrounds = 100, max_depth = 20 is", Specificity*100, "%.\n")
cat("The F score of model XGB with SMOTE, nrounds = 100, max_depth = 20 is", F_score*100, "%.\n")
## The Precision of model XGB with SMOTE, nrounds = 100, max_depth = 20 is 53.38346 %.
## The Sensitivity of model XGB with SMOTE, nrounds = 100, max_depth = 20 is 57.72358 %.
## The Specificity of model XGB with SMOTE, nrounds = 100, max_depth = 20 is 87.0021 %.
## The F score of model XGB with SMOTE, nrounds = 100, max_depth = 20 is 55.46875 %.

Summarize Running Time

Prediction performance matters, so does the running times for constructing features and for training the model, especially when the computation resource is limited.

Reweighting Sample

cat("Time for constructing training features=", tm_feature_train[1], "s \n")
cat("Time for constructing testing features=", tm_feature_test[1], "s \n")
cat("Time for training model=", tm_train_rw[1], "s \n") 
cat("Time for testing model=", tm_test_rw[1], "s \n")
## Time for constructing training features= 0.876 s 
## Time for constructing testing features= 0.187 s 
## Time for training model= 219.993 s 
## Time for testing model= 0.246 s

SMOTE

cat("Time for constructing training features=", tm_feature_train[1], "s \n")
cat("Time for constructing testing features=", tm_feature_test[1], "s \n")
cat("Time for training model=", tm_train_sm[1], "s \n") 
cat("Time for testing model=", tm_test_sm[1], "s \n")
## Time for constructing training features= 0.876 s 
## Time for constructing testing features= 0.187 s 
## Time for training model= 272.896 s 
## Time for testing model= 0.256 s